Comment by lol768 | Hacker News Reader

Comment by lol768 | Hacker News Reader

This could work great if the assistants actually had the full knowledge you'd expect from a human answering the phone. Usually if I've resorted to calling though, it's because I can't find out the answer from public sources. For example, "Can you accommodate this dietary requirement?", "Are you able to make dish X without ingredient Y?", "What sort of time of the day/week would be best for a romantic meal when you're not too busy?", "Are there any dishes you're out of, if we swing by in an hour or two?"

You can bet that these virtual assistants won't be equipped with enough knowledge, though I was impressed it knew there wasn't any outdoor seating (example in article).

Full knowledge is easy enough for 90% success. Just record every human to human conversation for a month and use a RAG or fine tune the model on the transcriptions.

Works okay until policies or other variables change. Then it's just confidently providing bad information which results in additional support time. And when it's wrong/lying 10% of the time what's the liability impact - what if the question is about handling allergens in food? Also updating that kind of knowledge base is going to be a nightmare.

> "Are there any dishes you're out of, if we swing by in an hour or two?"

> Full knowledge is easy enough for 90% success. Just record every human to human conversation for a month and use a RAG or fine tune the model on the transcriptions.

How could it possibly work that training on historical conversations would answer whether we are out of chicken today?

Well that’s the 10% it cannot answer. Most questions are much more mundane.

All of the mundane questions can be answered by going to the homepage.

The non-mundane questions are the ones you call for.

This is a solved problem for every customer-facing AI already. In this case, the chefs update the system prompt each day, either directly, or through a simple CRUD app.

Reply on news.ycombinator.com